- Bot Eat Brain
- Everything we know about OpenAI's Q* (Q-Star)
Everything we know about OpenAI's Q* (Q-Star)
PLUS: Anthropic's Claude 2.1 got rizz
Good morning, human brains. Welcome back to your daily munch of AI news.
Here’s what’s on the menu today:
Reportedly, this is partly why Sam Altman got canned 🛑
Q*, OpenAI’s new model, is so powerful that it freaked the board out.
Don’t watch entire videos, ask Bard instead 😴 ⏰
Bard’s new YouTube extension analyzes and summarizes entire videos.
Anthropic’s newest AI model is here 🤓 💻
Claude 2.1 gets a larger token window, more truthful responses, and more.
Is this why Sam Altman was fired? 🛑
Last Monday, we reported on OpenAI firing Sam Altman, its CEO. The chief Technology Officer, Mira Murati, was appointed interim CEO. Greg Brockman, OpenAI’s President, immediately resigned as well.
Chaos ensued. 👿 🔥
Sam Altman, Greg Brockman, and 95% of OpenAI’s employees agreed to leave and work at an AI research unit at Microsoft.
Then, last Wednesday, we covered OpenAI’s announcement that Sam Altman was back as the CEO. It also stated that the board responsible for firing him was out, and a new board would take its place.
Did he get fired again?
Nope, nice try though.
Last Thursday, reports surfaced of an AI model called Q* (pronounced Q-Star). It’s allegedly one of the factors that led to Sam Altman’s firing.
Isn’t it Sam’s job to create AI models?
Yes, but this one was so powerful that it reportedly caused safety concerns among OpenAI’s staff. OpenAI's researchers alerted its board about Q*’s potential for threatening humanity.
What makes it so threatening?
Apparently, Q* shows an unprecedented ability to solve math problems it had not seen before at a grade-school level.
It also showed a fast pace of development that is considered a breakthrough in AGI (Artificial General Intelligence). This was enough to alarm OpenAI’s safety researchers.
They wrote a letter to the board warning about a powerful AI discovery that could be a threat to humanity.
It’s scary because it does simple math…?
While it might not seem that impressive, this is a first for AI models. Up till now, AI models like GPT-4 haven’t been trained on enough mathematical concepts to develop a deep understanding of them.
But GPT-4 helps me with math all the time.
It’s not that GPT-4 can’t solve any math equations at all, but the mathematical content in its training data is small compared to the large range of text it understands. (That’s why it’s called a large language model).
GPT-4’s mathematical limitations also come as a result of its sequential processing approach. This means it processes text one token at a time.
So, why would Sam Altman be fired over this?
That isn’t clear, and OpenAI hasn’t made any official comments about Q* yet.
FROM OUR PARTNERS
Meet startup founders making similar revenue as you.
Indie Worldwide is a global network of bootstrapped startup founders.
What you'll find at Indie Worldwide:
Passionate founders and indie hackers to bounce ideas off of.
Business partners, clients, and beta testers.
Motivation and support from people going through the same journey.
A global network impossible to create if you only meet people in your own city.
Join the club to get matched 1 on 1 with founders who have similar revenue as you, access $20,000+ in benefits for your startup, and connect with 500+ bootstrapped startup founders in our moderated Slack.
Don’t watch videos, chat with them 😴 ⏰
A couple of weeks ago, we reported on Google DeepMind’s “Levels of AGI.” It’s a framework that categorizes how close AI models are to AGI status.
Then, about a week later, we covered Google DeepMind’s GraphCast. It accurately generates 10-day global weather forecasts in less than a minute.
More Google DeepMind stuff?
Nope, I got you again. 🤣
Why did they bother?
Google claims the upgrade was in response to requests for deeper engagement with YouTube content.
Its goal is to make Bard more useful in parsing and extracting key video information.
How does it work?
Bard’s new YouTube extension allows you to ask specific questions about content within YouTube videos.
Early tests showed Bard to be quick and mostly accurate in summarizing video content and answering specific follow-up questions.
Why would I use this?
Google claims this opens up possibilities for Bard in educational contexts, multimedia search, podcast engagement, and summarization.
But basically, why watch a video all the way through when you could ask a question about it?
So people won’t watch my videos anymore?
Google didn’t say anything about how this would impact content creators’ compensation.
A LITTLE SOMETHING EXTRA
Claude 2.1’s got the rizz 🤓 💻
In September, we covered Anthropic and Amazon’s $4 billion collaboration. Amazon invested the money in exchange for 49% of Anthropic’s company.
Last week, Anthropic launched its latest AI model, Claude 2.1. It improves upon Claude 2 with a longer context window, more truthful responses, and more.
A bigger context window, you say?
You bet. Claude 2.1 can now understand up to 200,000 tokens.
This means you can upload documents up to 150,000 words (over 500 pages) and use Claude 2.1 to summarize it, analyze it, and more.
Sweet. What about its truthfulness?
Anthropic claims it cut its rate of making incorrect or misleading statements by 50%
It also added features to improve the reliability and truthfulness of its responses.
Anthropic also added a new beta feature that allows you to integrate Claude 2.1 with databases, APIs and more.
This means you can use Claude to assist you in more of your day-to-day tasks. You can have it use a calculator for complex numerical processing, answer questions by searching databases, and more.
The OpenAI drama turned me off, how do I switch to Claude 2.1?
It is now powering Anthropic’s chat interface for both the free and Pro tiers.
You can also use it in Anthropic’s API. Keep in mind, the 200,00 token capability is only accessible to its paying customers.
MEMES FOR DESSERT
YOUR DAILY MUNCH
Trace AI — an AI tool to easily build iOS apps with text-to-UI code capabilities.
Gordon RamsAI — upload a picture of your food and have Gordon roast it.
OnlyPans — AI-powered meal prep, recipes, and schedule tool.
Mojju — a library of custom GPTs that you can integrate with Zapier and more.
An entertaining take on OpenAI’s Q*. Is this blown out of context, or something to really worry about?
What is OpenAI’s game plan? How Sam Altman, and OpenAI as a whole, has handled this transition strategically.
Andrej Karpathy’s intro to LLMs. (Video).
NVIDIA delayed its H20 AI chip. The impact of the United States’ recent trade restrictions on China is a factor in this decision.
Google open-sourced Project Guideline. It uses AI to guide visually impaired users along outdoor paths.
Google Meet introduced new hand gesture detection features. The point is to allegedly enhance communication efficiency.
FusionFrames — a text-to-video generation architecture that achieves top scores in many different evaluations.
GAIA — Meta, HuggingFace, AutoGPT, and GenAI’s benchmark to evaluate AI assistants in real-world reasoning and multi-modality tasks.
PG-Video-LLaVA — the 1st LLM with pixel-level grounding for videos that integrates audio to enhance understanding.
TWEET OF THE DAY
Ethan Mollick, a renowned Wharton University Professor, posts about GPT-4’s ability to scan screenshots to find deals.
Tag us on Twitter @BotEatBrain for a chance to be featured here tomorrow.
“Sunset Above Mountain Tops“ by @KlatuBaradaNiko
Until next time 🤖😋🧠
What'd you think of today's newsletter?